AMD: GPGPU Accelerators in CPU Sockets Make No Sense [UPDATED]

AMD Proclaims Torrenza Platform Dead

by Anton Shilov
12/11/2011 | 06:08 PM

UPDATE: Clarifying position and technology implementation by Intel Corp.

 

Advanced Micro Devices said that making special-purpose GPU-based accelerators compatible with CPU sockets makes no sense. The approach proposed by AMD is vastly different from the one that is allegedly offered by its arch-rival, Intel Corp., which recently demonstrated its Knights Corner accelerator in an LGA form-factor that is used for microprocessors.

In a conversation with X-bit labs' Anna Filatova, AMD's leading software expert Neal Robison said that Fusion-architecture - which integrates general-purpose [x86] processing cores with highly-parallel stream processors of Radeon GPUs - is a better solution for high-performance computing than to install special-purpose accelerators into CPU sockets. According to AMD, "it makes more sense from the software developers standpoint". Besides, it investments into "tool has already been made so we might as well use it". It looks like the once proposed Torrenza platform is no longer even considered as viable.

"APU is a better and cleaner solution than sticking a GPU in the same socket," said Neal Robison.

 

AMD once - in the early 2000s - proposed a solution that allowed special-purpose accelerators to be installed into the same sockets as AMD Opteron microprocessors for servers. Although no HPC supercomputer projects with Torrenza have been launched, the idea continues to live on. Intel, based on certain assumptions, wants its Knights Ferry highly-parallel accelerator to be installed into sockets for Xeon server chips.

While graphics processing units (GPUs) provide extremely high theoretical compute performance, their speed is limited by a number of factors. The main factor is software that is not efficient enough to take all the advantage that the GPU has to offer and also cannot use features already available in x86 microprocessors; another factor is PCI Express that connects CPUs and GPUs and has bandwidth and latency limitations; yet another thing is performance per watt as GPUs offer highly improved performance per watt compared to GPUs. The accelerated processing units, or APUs, majorly strike out every drawback of CPU+GPU structure except software factor.

The actual model of high-performance computing as well as typical PCs involves numerous technologies, including those with interconnection. Only time will tell which will be the most efficient one.